OpenAI Blog
·
3 weeks ago
Samsung Electronics deployed ChatGPT Enterprise and Codex across its global workforce, making it one of OpenAI's largest corporate implementations. The company integrated these tools to support employees in customer service, software development, and internal operations across multiple departments. Samsung employees now have access to generative AI capabilities for routine business tasks, potentially reducing processing time for common inquiries and code generation work.
TheSequence
·
3 weeks ago
SpaceX agreed to acquire Cursor, an AI coding editor, for $60 billion in stock to strengthen its xAI division, while top AI researchers including Noam Shazeer and John Jumper departed Google for OpenAI and Anthropic respectively. Google's $2.7 billion acqui-hire of Shazeer two years prior failed to ensure retention, demonstrating that researcher talent rather than compute capacity has become the limiting factor in AI development. AI capabilities are expanding beyond software into hardware and biology, with Midjourney building a full-body ultrasonic medical scanner and various startups raising hundreds of millions for embodied AI and world models, signaling that AI infrastructure has become strategically essential across physical industries.
Exponential View
·
3 weeks ago
A Harvard Business School paper finds that AI-native startups are 25% smaller than comparable non-AI startups but have more engineers and denser expertise, with real value from AI coming when it is embedded directly into the product rather than added as a tool. Generalist frontier models are outperforming specialized medical AI tools in head-to-head testing, contrary to earlier expectations. The results support Richard Sutton's 2019 thesis that general computational methods ultimately outperform purpose-built specialized approaches across domains.
Eugene Yan
·
3 weeks ago
Researchers have developed patterns and benchmarks for evaluating whether AI models can find and exploit security vulnerabilities, including Cybench (40 capture-the-flag tasks with difficulty ranging from 2 minutes to 25 hours), CVE-Bench (40 critical real-world vulnerabilities), and CyberGym (1,507 memory-safety bug instances). The best-performing models achieved only 17.5% success on unguided Cybench tasks and 12.5% on CVE-Bench zero-day scenarios, with all models failing on challenges above 11 minutes difficulty. These evaluation frameworks enable security teams to measure when AI agents become useful for defense versus when they risk assisting attackers, using sandboxed environments, grading mechanisms, and partial-credit scoring along attack chains.
